AI startups take on risk as a business model

We've discussed at length about how data and AI can take software products to the next level, so we'll switch gears this time to explore ways it enables new business models altogether. Data-driven businesses track the who, what, when, and where of a process. This allows you to get a clearer idea of why something happens and informs you of howto control the four Ws as inputs to steer toward the target outcome. When you're able to not only accurately predict an outcome but also control your production to steer toward that desired outcome, you gain intelligence and operational leverage: this is an unprecedented opportunity to control your own destiny while taking on risk. Here's an overview of models I'm seeing take shape among companies building businesses around AI:

The bet of AI businesses

The fundamental business "bet" you're making in building an AI model is, once you have sufficient data, the model will recommend more accurate or efficient decisions than you would be achieve on your own. AI startups will be hard pressed to find customers who will be willing to take that performance risk on them as an unproven vendor, however, so many of them take on that risk themselves. This becomes the familiar AI startup strategy of bearing the risk and cost (and margin erosion) of employing humans in the loop to simultaneously train the model while filling in for performance gaps, in order to deliver a functioning recommendation system to the customer and access training data.

Pairing insight with action

We often conflate intelligence and automation, but the intelligence derived from a data science or AI model on its own only provides recommendations or insights. Recommendations are only as valuable as what is done with them, and many companies selling AI solutions will find that their customers are either not equipped to, or not incentivized to, utilize the recommendations to achieve the fullest extent of the efficiency they could achieve. This means that, even if you succeed in getting your AI model into the hands of customers, they may not renew after failing to achieve meaningful ROI.

The AI products that layer on automation to act on the recommendations, rather than only serving up insights, stand a greater chance of delivering meaningful value to customers and surviving the first renewal cycles. Unfortunately, it's rare to see customers give up partial control of a process. If the target workflow is considered high risk, such as deciding how to treat an emergency room patient, the customer will want to control the decision and only consume the AI as a recommendation tool that they can, and will, ignore if it makes an unexpected recommendation, dramatically curtailing the impact of the platform. If the target workflow is low risk, the customer will likely have outsourced it entirely to a vendor-- as is commonly the case for functions such as IT and security. In this case, these external partners are your true customers. Since their enterprise customers only care that the work is done reliably, not how it is done, these managed service providers historically under invest in technology and rely heavily on labor, leaving little operating budget to spend on software.

Closing the book on MVPs and "Lean"

The realities outlined above mean that optimizing your chance of building a successful business around an AI model and automation requires you to take on more start-up risk than ever by abstracting as much of the AI tech stack as possible and only deliver the finished product or completed service. This may seem to run counter to the classic Silicon Valley wisdom that identifying a minimum viable product (MVP) from which to expand iteratively is the most capital efficient way to go to market. That's because the principle behind "Lean Startup" and MVPs were only a means to the end goal of maximizing capital efficiency-- for SaaS startups, this was truly the least costly way to go to market. I've outlined at length -- and others are starting to see the same thing -- that AI software is more expensive to bring to market than traditional software. That's because, in order to deliver ROI, you not only need to integrate into the system of records to access data for your models, you also have to connect to the systems of control in order to put those insights to work and achieve that potential efficiency gain. When every customer has an idiosyncratic combination of systems of record and control, deployment becomes exponentially more complex than traditional software. Operating the workflow on behalf of your customer means you only need to build one custom integration at the input, because you only have to consider your own system of control.

When you take on the risk of owning the workflow, you control and can optimize the means of production to utilize the AI to its fullest extent. You also drastically simplify adoption for the customer who may only have to give you the raw inputs or dictate a specific output, without having to invest in the process digitization or data infrastructure required to utilize an AI model themselves. While VCs have historically abhorred these managed services models because they are difficult to scale, AI-driven automation allows you to bet that as you collect more data to improve your model, it will require fewer corrections, freeing up capacity for you to take on more customers without having to increase headcount. By taking on more risk at the outset, you access more upside as the model increases in accuracy, allowing you to keep the efficiency gain as profit.

Risk as a business model

Owning the means of production, combined with your ability to predict outcomes with increasing accuracy, positions you to grease the skids to accelerate adoption and capture even more value than was ever possible from merely selling software. When you can predict outages or failure, you can guarantee a minimum level of performance and operate at tighter tolerances, reallocating as profit the margin you previously had to leave as buffer for error. This is a practice that is much higher risk for non AI model-driven managed service providers, who run on thin margins and may face highly variable input costs. As you become even more confident in your future performance, or as a way to sweeten the pot by reducing downside risk of adoption, you could even charge customers a nominal baseline fee to partially cover your operating costs and make up the difference by taking a share of the upside you generate for them-- in exchange for strategically taking on the customer's risk of adoption, you can capture even more value in the end.

Now that the initial AI hype has subsided, so, too, has the seemingly limitless willingness to indiscriminately try AI solutions. The companies that find the right business model to wield risk to their advantage and maximize value capture will build the next era's breakaway tech companies. In the interest of confidentiality-- and because many of them are still figuring out how to apply the models discussed above-- I didn't go into specific detail here. If you're interested in whiteboarding what this might mean for your own ideas, drop me a line -- don't be shy!

Recent Posts

See All


Thanks for submitting!


©2018 by AI ROI TOOLKIT. Proudly created with